Low-Rank Matrix Completion [Lecture Notes

Abstract

Imagine one observes a small subset of entries in a large matrix and aims to recover the entire matrix. Without a priori knowledge of the matrix, this problem is highly ill-posed. Fortunately, data matrices often exhibit low-dimensional structures that can be used effectively to regularize the solution space. The celebrated effectiveness of principal component analysis (PCA) in science and engineering suggests that most variability of real-world data can be accounted for by projecting the data onto a few directions known as the principal components. Correspondingly, the data matrix can be modeled as a low-rank matrix, at least approximately. Is it possible to complete a partially observed matrix if its rank, i.e., its maximum number of linearly independent row or column vectors, is small?